Diffusion of real versus misinformation during a crisis event: A big data-driven approach

Highlights

• Tweets that were novel were more likely to go viral.

• Tweets with negative sentiment diffused faster than neutral or positive ones.

• The lower the lexical density, the higher the number of retweets.

• Association between novelty and retweet count is the same for real vs. misinformation.

Abstract

Misinformation has captured the interest of academia in recent years with several studies looking at the topic broadly with inconsistent results. In this research, we attempt to bridge the gap in the literature by examining the impacts of user-, time-, and content-based characteristics that affect the virality of real versus misinformation during a crisis event. Using a big data-driven approach, we collected over 42 million tweets during Hurricane Harvey and obtained 3589 original verified real or false tweets by cross-checking with fact-checking websites and a relevant federal agency. Our results show that virality is higher for misinformation, novel tweets, and tweets with negative sentiment or lower lexical density. In addition, we reveal the opposite impacts of sentiment on the virality of real news versus misinformation. We also find that tweets on the environment are less likely to go viral than the baseline religious news, while real social news tweets are more likely to go viral than misinformation on social news.

Introduction

Online social networks (OSN) have become an efficient means of information dissemination (Bowler, Halbesleben, Stodnick, Seevers, & Little, 2009). For example, statistics show that in 2017, 67% of US adults depended on online social network platforms such as Twitter, Facebook, and Snapchat for news, as compared to 62% in 2016 (Moon, 2017). Close to 9 out of 10 Twitter users rely on the social media outlet primarily for news, and of those who do, 74% use the platform daily (Rosenstiel, Sonderman, Locker, Ivancin, & Kjarval, 2015). As a result, Twitter has come to replace mainstream media as the number one choice for news, especially among millennials and younger; its reach is expected to grow even further (Stieglitz & Dang-Xuan, 2013). With the advent of social media and portable communication devices, disseminating information during crisis events has become easier and more seamless due to the steady increase in technology adoption (Chaubey and Sahoo, 2021, Dwivedi et al., 2020, Palen et al., 2010, Roy et al., 2020, Shklovski et al., 2010, Shklovski et al., 2008, Sinha et al., 2019). For example, during natural disasters such as hurricanes, information is rapidly propagated through social media platforms using mobile devices and augmented by local eyewitness accounts (Oh, Kwon, & Rao, 2010). Despite its popularity, OSN as an information diffusion channel has an inherent disadvantage. Because users of these platforms act as gatekeepers and are prone to their own individual biases, it is difficult to assess the veracity of the news items being propagated. This issue has been exacerbated even more by the sheer amount of data that flows through online social networks. This phenomenon has led to both factual events and misinformation items on social media reaching a larger audience than news from major mainstream news outlets (Allcott & Gentzkow, 2017). Hence, it is not surprising that headlines of misinformation deceive American adults about 75% of the time, and the most popular misinformation stories usually garner far more shares than authentic news (Silverman & Singer-Vine, 2016).

Extant research has identified salient features that may not only influence the diffusion of information, but also help in assessing its veracity on social media (Boyd, Golder, & Lotan, 2010; Lee, Mahmud, Chen, Zhou, & Nichols, 2014). These features may be broadly grouped into two research themes: information diffusion and information veracity. Information diffusion research has identified three types of features, which include the following: (a) user-based features that directly relate to the behavior and characteristics of users; (b) time-based features related to the time a tweet is generated and posted; and (c) content-based features related to the contents of text embedded in the messages being propagated (Hoang & Mothe, 2018). Information veracity research, which focuses on more proactive measures like detection, relies on the following features: (a) linguistic cues related to the linguistic properties of the messages; and (b) social network characteristics, which predominantly rely on user profile characteristics that can be extracted from the metadata collected in the network (Conroy et al., 2015, Rubin et al., 2015). Both approaches rely on similar features (content- and user-related) but depend on different methods. Information diffusion studies involve several approaches, including a combination of sentiment and text analysis and feature extraction and engineering methods for both predictive and descriptive analysis. Conversely, information veracity predominantly involves the use of prescriptive analysis (Hoang and Mothe, 2018, Rubin et al., 2015, Shin et al., 2018, Suh et al., 2010).

Big data often collated from social media and other repositories provide a wealth of information for information systems (IS) researchers. However, such data come with their own challenges, primarily due to their high volume, veracity, variety, and velocity. This issue has made it increasingly difficult for researchers to not only effectively access the data and ensure its integrity, but also to analyze it. Given the complexity of analyzing Twitter data and the unavailability of complete datasets due to Twitter’s rate limits (https://developer.twitter.com/en/docs/twitter-api/rate-limits), large scale studies using both econometric and algorithmic techniques have also been rare. In addition, much of the extant big data-driven research focuses on describing what phenomena are happening rather than theory building to explain the causality of the observed events (e.g., (Kitchens, Dobolyi, Li, & Abbasi, 2018; Zhou et al., 2018)). Hence, in alignment with Kar and Dwivedi (2020)'s call for more theory building in big data-driven research, we examine the antecedents to tweet virality and how the impacts of these antecedents differ for authentic news vs. misinformation by using data collected from Twitter during a crisis event.

Our research contributes to explaining the inconsistent results from a handful of analyses using Twitter data. For example, a study using textual contents from Twitter showed that both positively and negatively charged tweets were retweeted more often and quicker than neutral ones (Stieglitz & Dang-Xuan, 2013). The researchers concluded that sentiments inferred from social media contents might be positively associated with information diffusion. Other studies have related the propagation of misinformation to automated entities and claimed that these entities actively spread misinformation. They showed that in the earlier phases, the automated entities target mostly influential users on social media that eventually lead to more diffusion of misinformation (Shao et al., 2018). This finding was contradicted by another study that found contrary to conventional wisdom, malicious entities were not any more responsible for the propagation of misinformation than humans (Vosoughi, Roy, & Aral, 2018). Rather, the study claimed that human behavior contributed more to the differential propagation of misinformation than automated entities.

The current study employs a unique dataset of tweets collected across a 5-week period during Hurricane Harvey to empirically investigate the virality of both real and misinformation during a crisis event based on features extracted. Specifically, we investigate the following:

• What time-based, content-based, and user-based factors affect the virality of authentic news versus misinformation on social media during crisis events?

• How are the impacts of time-based and content-based factors on virality different for authentic news versus misinformation during such events?

• How does the virality differ for authentic news versus misinformation in different news categories?

Our empirical results show that virality, measured by the retweet count, is higher for misinformation, novel tweets, and tweets with negative sentiments and those with low readability. In addition, the impacts of sentiment are different for misinformation and authentic news. Tweets on the environment have lower retweet counts compared with the baseline religious tweets, and the retweet counts of social tweets are higher for authentic news than misinformation. Despite the burgeoning literature on misinformation diffusion, this study is the first to address the readability of misinformation by examining its lexical component. Our results show that when a tweet contains more lexical words, users share it less. These findings have implications for research and practice and provide guidelines for administrators in online social networks.

The rest of this study is organized as follows: In the next section, we provide a review of the background literature on misinformation, information diffusion and virality, and information veracity. Next, we provide our conceptual model and develop a set of hypotheses in Section 3. We discuss our approach to data extraction and analysis in Section 4, followed by the results in Section 5 and a discussion of the theoretical contributions and practical implications in Section 6. We conclude by discussing the limitations and implications for future research.